Code Intel Pipeline

One real run against a fresh clone of expressjs/express (a3714473) — unedited output. 把陌生仓库摊成一张地图，再让 Agent 动手。

$ git clone https://github.com/expressjs/express.git
$ code-intel ./express

[PASS] ./express
  Outcome: completed
  Run evidence: <data-root>/artifacts/express/<run-id>/run-complete.json

213files inventoried

124symbols extracted

504imports resolved

7 / 7pipeline nodes passed

~15 swall clock, local, no network

Agent code map — where should an agent start reading?

Every file is scored by entrypoint, symbol, and import evidence (agent-code-slice-ranking.v1). Test, example, and fixture paths are tagged but never outrank the production sources they exercise. Top of the ranking:

score	file	evidence
45	index.js	entrypoint, imports
35	lib/response.js	sendfile, onaborted, ondirectory, onerror…
35	examples/view-locals/index.js	demoted: example dir — no entrypoint bonus
35	test/app.router.js	tagged test — no rank bonus
30	lib/utils.js	acceptParams, createETagGenerator, parseExtendedQuery…
30	lib/application.js	symbols, imports

What one run commits

repo.snapshot
inventory.rg
evidence.native-code
evidence.graph
evidence.sentrux
diagnosis.hospital
doctor

Artifacts land in a content-addressed store and are committed atomically — a run is either fully published or not indexed at all. Read it back with code-intel resume --repo ./express.

Governance verdict

# Code Intel Hospital Report

- Status: green
- Disposition: observe
- Primary diagnosis: clean snapshot
- Next protocol: post_op

## Treatment
- Keep this admitted evidence set as the clean comparison baseline.

On repositories with structural damage the same node emits a surgery plan; Sentrux rule and baseline gates fail the run on structural regression instead of letting an agent pile speed onto debt.

Reproduce it

$ git clone https://github.com/expressjs/express.git   # full history; unshallow if --depth 1
$ code-intel ./express
$ code-intel resume --repo ./express --json